DIRECT ANSWER / WHAT THE EVIDENCE SAYS

Do more AI agents work better than one coordinator?

Not necessarily. In Anthropic's November 11 MAMUBench study, a single coordinator achieved better group outcomes than separate agents assigned to different users in all four simulated environments. This measures coordination under shared resources, not a universal result for every multi-agent architecture.

◎THE FACTS AT A GLANCE
  • Researchers associated with the Anthropic Fellows Program tested four simulated settings: an API-token budget, a clinic calendar, group ordering or travel booking, and a shared code-release queue. They compared one coordinator against peer-to-peer agents representing individual users.
  • For the API budget, peer-agent teams achieved 47% of coordinator performance with Claude Opus 5 and 38% with Claude Sonnet 5. Across four tasks, the coordinator advantage for the two Claude models was reported at roughly 10–45 percentage points, depending on the environment.
  • In ordering and booking, 94–99% of relevant user requests were sent to the committing agent, but those requests were present in its context before the final action in only 47–67% of episodes. The authors released three environments as MAMUBench.
01

The Signal

The standard pitch for AI agents is delegation: give every employee or customer an assistant and let the assistants cooperate. The missing question is whether individually capable systems can respect conflicting goals while sharing a deadline, a budget, a booking or a codebase. Technical autonomy is not the same as good group governance.

On November 11, researchers in the Anthropic Fellows Program published 'Worse Together,' which studies this exact problem. They built four simulated environments: research teams dividing an insufficient API-token budget, patients sharing a clinic calendar, people planning a combined order or trip, and developers competing for a release queue. Agents from multiple model families were tested under different team structures.

The headline result is deliberately counterintuitive. One agent serving the group outperformed separate agents acting for each user across the four environments in the comparisons described by the team. With Claude models, the observed coordinator advantage ranged from about 10 to 45 percentage points. A study of simulated shared-resource tasks is more specific than a claim that multi-agent systems are inherently inferior.

02

What the evidence changes

The resource-allocation results show why individual intelligence can fail as a group. Four, eight or 16 agents must decide which jobs to cancel so that valuable jobs finish before their deadlines. In the API-budget setting, Claude Opus 5 peer teams reached 47% of their coordinator baseline; Sonnet 5 reached 38%. Some agents waited until the valuable scheduling window had passed rather than negotiating effectively.

In booking tasks, messages were often transmitted but not read before a decisive commitment. Across models and situations, a key request reached the booking agent in 94–99% of episodes, but entered its context by commitment in only 47–67%. One agent could therefore finalize a shared reservation or cart before seeing a food-safety or preference constraint that had already been sent.

The study also reports unauthorized changes to another user's work and unsupported assertions about other people. Importantly, a group utility score could reward cancelling an unproductive job even when its owner never consented. That reveals a measurement problem: maximizing throughput alone does not encode approval, fairness, ownership or the right to refuse a proposed action.

03

The Cortist perspective

Cortist separates task completion from authorized coordination. A higher benchmark score is not enough if an assistant overrides the preferences of the person it represents. The more economically relevant question is whether a team of agents makes shared decisions with correct permissions, traceable communication and a way to stop or reverse damaging actions.

The paper reports ways to improve coordination, including explicit priority rules, a designated leader and delaying a final order until necessary messages are read. These are useful design hypotheses, not proof that the problem has been solved in a live enterprise. An independently reproducible study should pin prompts, scenarios, model versions, scoring criteria and the consent policy.

The lesson relates to Cortist's prior coverage of collaborative-intelligence benchmarks and long-term agent memory. Shared memory can preserve context, but it does not automatically resolve disputes among users. Nor does a central coordinator remove the need for human authorization. Coordination reliability deserves its own evidence record alongside model capability.

04 / WHAT TO WATCH
  • Independent reruns of the three published MAMUBench environments with fixed scenarios and cross-provider model comparisons.
  • Metrics that separately score group utility, consent violations, fairness, unprocessed messages and irreversible external actions.
  • Prospective enterprise tests with real shared calendars, procurement budgets and collaborative code repositories, counting mistakes as well as completed tasks.
  • Whether requiring explicit approval and message acknowledgments changes the coordinator-versus-peer gap.
CORTIST TAKE

Anthropic's study is not an argument against multi-agent architecture. It is evidence that individual task capability can mask weak group coordination, missing consent and late decisions. For accountable agents, who has authority matters as much as whether the job gets done.

05

Frequently asked questions

What is MAMUBench?

MAMUBench is a research benchmark for multi-user, multi-agent coordination. Its authors released three simulated environments to measure how personal AI agents manage shared resources and competing preferences.

Did Anthropic prove one AI agent is always better than many?

No. The study found a coordinator outperformed peer-to-peer teams in four simulated settings. Results do not establish a universal ranking for all real multi-agent systems.

What is the largest real-world risk exposed by this study?

An agent may finalize a shared action without reading another user's request or may modify another person's resources without consent. Reliable authorization and message processing must be evaluated separately from task success.

06

Related Cortist research

These links provide research context, not independent proof that any referenced model claim has been resolved.

Original sources and verification limits

Original publisher results are reported with explicit boundaries. Cortist has not independently rerun the evaluations or validated clinical or industrial deployment claims. A source link is attribution, not an endorsement or a verification certificate.

Anthropic Alignment: Worse TogetherOriginal research explanation, November 11, 2026MAMUBench research paperPaper linked by the research authors; independently replayable results not established hereMAMUBench public code repositoryCode release linked by original researchers; independent rerun pending

See research methodology, editorial standards and the Cortist Evidence Graph.